Communications of the ACM
Next-Generation Supercomputers
Supercomputing performance upgrades are unlikely to be as spectacular in the next decade as they were in the last two, writes University of Notre Dame professor Peter Kogge. The U.S. Defense Advanced Research Projects Agency hoped that an exaflops-class supercomputer would be practically realizable by 2015, but a panel Kogge organized to debate this question concluded that such a breakthrough requires a complete rethinking of supercomputer construction in order to dramatically minimize power consumption.
An additional challenge is keeping a massive number of microprocessor cores busy at the same time. Kogge says that "unless memory technologies emerge that have greater densities at the same or lower power levels than we assumed, any exaflops-capable supercomputer that we sketch out now will be memory starved."
Another daunting challenge is providing long-term storage with sufficient speed and density to retain checkpoint files, while reducing the operating voltage would make the transistors susceptible to new and more frequent faults.
Nevertheless, Kogge thinks exaflop systems are attainable, but creating such a supercomputer "will demand a coordinated cross-disciplinary effort carried out over a decade or more, during which time device engineers and computer designers will have to work together to find the right combination of processing circuitry, memory structures, and communications conduits."
From IEEE Spectrum
View Full Article
No entries found